Cohere North 2: Agent Platform Adds Token Spending Caps

Cohere has released North 2, a major update to its platform for running AI agents inside large organizations. The release centers on two changes: a rebuilt system for orchestrating multistep agent work, and new controls that let administrators cap how many tokens their teams spend.

The pitch is aimed at a familiar enterprise problem. According to Cohere, companies deploying agents have usually had to give up one of three things - security, capability or cost control - to get the other two. North 2 is the company's attempt to remove that trade-off.

North has been generally available since August 2025. Cohere says a year of large production deployments, in sectors ranging from finance to the public sector, shaped the upgrade.

How the new orchestration works

The redesigned orchestration system handles multistep tasks on its own. When a step needs approval, it brings in a human. A few other additions build on that:

• Shared agents. An agent built from a simple prompt can be shared across the company. Other teams can reuse it instead of rebuilding the same workflow.
• Persistent memory. Each agent keeps its memory from one session to the next.
• Skills. Logic that agents use repeatedly can be packaged as a reusable skill.
• Shared libraries. Documents and other assets that every agent needs go into common libraries.

North 2 can also turn a natural-language request into a slide deck, a dashboard or a lightweight app, which users can prototype directly in the chat window. For recurring processes, Cohere offers prebuilt templates and a drag-and-drop workflow builder that reports performance in real time.

The connector library now includes Slack, GitHub, and Microsoft's SharePoint and OneDrive. Cohere plans to add connectors to financial data providers, including PitchBook Data, S&P Global and FactSet Research Systems.

Spending caps in North Admin

The cost controls live in North Admin, the platform's management console. It now tracks token use down to the individual user and the individual agent. Administrators can set consumption tiers for users and groups based on request and token rates, and cap usage across the whole company. Alerts can warn teams before they hit a limit.

North 2 is model-agnostic. That means the models an administrator assigns to each team can include models a customer brings from outside Cohere.

Cohere is also leaning on its work with Nvidia to lower running costs. The company says its models produce more tokens per second per node on Nvidia's Blackwell and Hopper GPUs. Kari Briski, Nvidia's vice president of generative AI, said this helps North 2 "deliver more tokens for less."

Security and deployment options

Security settings sit in North Admin as well. Guardrails can be applied agent by agent, screening prompts and responses for personally identifiable information and for attempted prompt injection. Autonomy policies restrict each agent to the actions it is authorized to perform.

Deployment options remain unchanged: North runs on-premises or in a customer's own virtual private cloud. That includes fully air-gapped installations for organizations that permit no outside connection at all.

Who is using it

Bell Cyber, a unit of Bell Canada, runs North inside its own security operations. Jawed Ahmad, its chief technology and AI officer, said North lets the unit run agentic AI with the control and sovereignty requirements its customers demand "built in from the start."

South Korean IT services company LG CNS uses North internally and offers it to customers across the country. Radical Ventures, an early Cohere investor, built an automation on North that prepares the briefing its team reads before calls with founders. Partner Parasvil Patel said the team previously spent hours gathering basic context.

The launch comes less than three weeks after Cohere signed a merger agreement with German AI developer Aleph Alpha. The combined company will keep the Cohere name.

Our Take

The most telling part of North 2 is not the orchestration, but the spending controls. Per-user and per-agent token tracking suggests that cost has become a real deployment blocker, not an afterthought. That fits a broader pattern we have covered, where business AI use rises while spending falls and buyers push for tighter budget discipline.

The security side follows a similar logic. Guardrails and autonomy policies that sit around the model, rather than inside it, echo efforts like IBM and CoreWeave's agent workload controls. Whether Cohere has actually removed the security-capability-cost trade-off is a claim customers will need to test for themselves. It is also worth watching how the Aleph Alpha merger shapes North, especially for European buyers focused on sovereignty.